Views: 205 Downloads: 23

Abstract

Accurate estimation of disease-specific survival is central to precision oncology because treatment decisions increasingly depend on individualized risk rather than broad population averages. This study developed and compared traditional and artificial intelligence-based survival models for predicting 5-year disease-specific survival among patients with non-small cell lung cancer. The analysis integrated clinical, genomic, and treatment-related variables to evaluate whether modern survival learning methods improve individualized prognostic estimation. The study cohort included 369 patients diagnosed with non-small cell lung cancer and followed for disease-specific survival over a 5-year horizon. Candidate predictors included tumor stage, age, Eastern Cooperative Oncology Group performance status, histological subtype, smoking history, EGFR mutation status, KRAS mutation status, tumor mutational burden, and treatment modality. Three survival models were estimated and compared: Cox proportional hazards regression, random survival forest, and DeepSurv. Model performance was assessed using time-dependent C-index, Brier score, and calibration across clinically relevant time horizons. DeepSurv achieved the highest discrimination, with a 5-year time-dependent C-index of 0.78, outperforming Cox regression and random survival forest. Calibration analysis indicated that DeepSurv produced predicted survival probabilities that were closely aligned with observed disease-specific survival across most risk strata. Explainability analysis was conducted using SHAP values adapted for survival outcomes to identify the most influential predictors in the best-performing model. Tumor mutational burden, performance status, and tumor stage emerged as the strongest predictors of 5-year disease-specific survival. Personalized risk curves demonstrated that patients with similar tumor stages could have substantially different predicted survival trajectories when genomic and functional status variables were jointly considered. The findings suggest that deep survival models can add measurable prognostic value in precision oncology when combined with interpretable model explanations. Although the retrospective design, relatively small sample size, and lack of external validation limit immediate clinical translation, the study provides empirical evidence that AI-based survival prediction can support individualized prognosis in non-small cell lung cancer. Future work should prioritize prospective validation, multi-center transportability assessment, and integration into decision-support workflows‎.

Keywords: Survival analysis, Artificial intelligence, Personalized risk prediction, Precision medicine, DeepSurv, Random survival forest

Citation Formats: